Weekly Meeting 2026‐01‐16

Weekly Meeting - 2025-01-16

🏅 Issue of the Week

[Enhancement] The design of the base class Kernel is unreasonable, and its formatting needs to be corrected

Issues:

To be discussed:

Development and testing tool chains:

• [Enhancement] Replace CSV-Based Benchmark Configuration with Pytest-Native Approach

• [Enhancement] Adopt a Modern Python Development Toolchain

• [Enhancement] Adopt pytest for Unit Testing

Technical problems:

• [Feature Request] Kernel workspace interface for higher-level APIs

• [Feature Request] Kernel cache

Assigned

To be solved:

MHA @YuchongLi

• [Feature Request] Multi-Head-Attention - Benchmark

CI @YuchongLi

• [Enhancement] Add Docker Images in CI
• [CI][Enhancement] Miss the tileOps install and build process
• [CI][Enhancement] Refactor CI functionality to enhance CI stability and expand test coverage
• [Enhancement] Move imports of operator baseline dependency libraries inside the baseline function

NSA @JienengYu

• NSA_fwd
• nsa_topk

MLA decode with kvcache @ZhejinXu

• [Feature Request] MultiHeadLatentAttentionDecodeWithKVCache Kernels
• [Feature Request] MultiHeadLatentAttentionDecodeWithKVCache Function
• [Feature Request] MultiHeadLatentAttentionDecodeWithKVCache Function
• [Feature Request] MultiHeadLatentAttentionDecodeWithKVCache Layer

manifold-constrained hyper-connection @AngGao

• [New Op] manifold Hyper Connection (mHC)

Closed this week:

GEMM, GQA, and MHA @YuchongLi

• [Refactor] Gemm - Code Format
• [Refactor] Multi-Head-Attention - Code Format
• [Refactor] Group Query Attention - Code Format

CI @YuchongLi

• [BUG] The CI configuration file fails to link with the appropriate CUDA compiler version

NSA @JienengYu

• refactor code format
• mean pooling

DeepSeek Sparse Attention Decode @YuxianDu

• [feature Request] Implement Indexer Kernel for DSA Attention in TileOps
• [Refactor] DeepSeek Sparse Attention Decode - Code Format

MHA Decode & GQA Decode @AngGao

• [Bug] The MHA_decode and GQA_decode kernels raise errors when a data type other than the default is specified.
• [Bug] The MHA_decode and GQA_decode kernels do not support sequence lengths that are either less than 128 or not a power of two.
• [Bug] Seqlen should not be in the init parameters for *_attention_decode
• [Bug] Benchmark program incapable in profiling GQA_decode when the sequence length

Grouped GEMM @QihangZheng

• [New Op Sub-task] Grouped GEMM - Code Format
• [Feature Request][GroupedGEMM] There is no profile implementation
• [Formatting][GroupedGemm] Pre-commit requires ruff for linting, but it misses ruff's configuration

PRs:

Merged:

• [BugFix] Support varying length and assigning data type in MHA_decode and GQA_decode kernels @AngGao

• [Feat]: Mean pooling @JienengYu

• [Refactor] DeepSeek Sparse Attention Code Reformat @YuxianDu

• [CI][Enhancement]Conduct daily testing on the main branch @YuchongLi

• [Refactor] Refactor MultiHeadAttention and GroupQueryAttention code format @YuchongLi

• [Formatting][GroupedGEMM] refactor the code organization and code format @QihangZheng

Action Items

For the next task, please try running each other’s examples.

Check whether they run correctly, identify any bugs, and note anything that feels inconvenient or unclear to use. If you have suggestions for new features or possible improvements, please share them as well.

Please record all issues, feedback, and suggestions in GitHub issues.

• gemm, @michaelwithu
• grouped gemm, @xZacky
• flash attention, @zhen8838
• flash decoding, @RMLYC
• deepseek nsa, @stelladuyx @RMLYC
• deepseek mla, @superAngGao